cleanup efa installer archive before install #6870
Conversation
|
Sorry - why was this closed? |
|
Hi could you please reopen this? We just spent a ton of money (and many hours) bringing up clusters with broken efa because a subset of nodes didn't have efa. Why? Because of this issue: Error downloading packages:
gcc-7.3.1-17.amzn2.x86_64: Insufficient space in download directory /var/cache/yum/x86_64/2/amzn2-core/packages
* free 11 M
* needed 22 M
Error: Failed to install packages.
Error: failed to install EFA packages, exiting
/var/lib/cloud/instances/i-08e860474f7d2683a/boothooks/part-001: line 7: pop: command not found
/usr/bin/cloud-init-per: line 63: /opt/amazon/efa/bin/fi_info: No such file or directoryThe archive needs to be cleaned up. This can't keep happening. I opened this over a year ago and I don't understand why it's been ignored and closed. What do you need from me? |
cPu1
added
priority/important-soon
|
@vsoch, please give us some time, we'll prioritize reviewing this PR. |
|
Thank you! Much appreciated. I looked at the code and I think the change needs to be added to two additional files - I'll do that shortly. |
Currently, the UserData section that runs during cloud init happens before any root volumes are expanded with growpart. Although the best solution would be to ensure the filesystem resize happens before these scripts are run, a quick means to fix the current issue is simply to cleanup the efa installer tar.gz, which is very large. I have tested this with hpc7g for a size 2 and size 8 cluster (previously both not working) and can confirm the devices are functioning after. Signed-off-by: vsoch <[email protected]>
vsoch
force-pushed
the
cleanup-efa-installer-archive
branch
from
747bf6d
to
d4eb786
Compare
|
Updated to include the same 2023 files. I also tested this today (this evening) and it fixed the issue I posted above - my cluster came up with all efa nodes. I'll need to try the experiments for the two clusters that failed tonight, but with less funds now, tomorrow. Thanks for the help and TBA speedy review! |
Problem: the node consistently runs out of disk space when adding efa, resulting in an unusable cluster with scattered nodes where the installer failed. Solution: the installer archive itself is huge, and we can simply remove it and avoid this error. Signed-off-by: vsoch <[email protected]>
vsoch
force-pushed
the
cleanup-efa-installer-archive
branch
from
d4eb786
to
a47a32a
Compare
|
Hi @cPu1 your bot closed the PR again! |
Sorry about that, I am not in favor of having this stale bot. We'll discuss this more. As for the PR, we will try to get this reviewed and released by next week. |
|
Thank you! And no worries about stale bot - it can be very helpful. I'm subscribed to the thread and am good to ping when it needs to be reopened. |
Currently, the UserData section that runs during cloud init happens before any root volumes are expanded with growpart. Although the best solution would be to ensure the filesystem resize happens before these scripts are run, a quick means to fix the current issue is simply to cleanup the efa installer tar.gz, which is very large. I have tested this with hpc7g for a size 2 and size 8 cluster (previously both not working) and can confirm the devices are functioning after.
And logs for a running node (what they should look like!)
This will close #6869
Checklist
README.md, or theuserdocsdirectory)area/nodegroup) and kind (e.g.kind/improvement)BONUS POINTS checklist: complete for good vibes and maybe prizes?! 🤯